Continual learning in medical devices: FDA's action plan and beyond

Artificial intelligence (AI) and machine learning (ML) software have the potential to improve patient care. An underlying algorithm can either be locked so that its function does not change, or adaptive, in which AI ML system performs continual learning. Continual learning, also known as lifelong is a technique decision logic of mathematical models updated through new data while retaining previously learned knowledge.1Parisi GI Kemker R Part JL Kanan C Wermter S with neural networks: review.Neural Netw. 2019; 113: 54-71Crossref PubMed Scopus (652) Google Scholar, 2Lee CS Lee AY Clinical applications learning.Lancet Digit Health. 2020; 2: e279-e281Summary Full Text PDF (43) Scholar By contrast, systems prevent ability learn from post approval, real-world data, thus cannot over time same way adaptive systems. has long been part computer science.1Parisi Yet, no medical device based on yet approved by US Food Drug Administration (FDA).2Lee 3Rivera SC Liu X Chan A-W Denniston AK Calvert MJ Guidelines for clinical trial protocols interventions involving artificial intelligence: SPIRIT-AI extension.BMJ. 370m3210Crossref (53) It likely FDA will make about such near future.4FDAArtificial intelligence/machine (AI/ML)-based (SaMD) action plan.https://www.fda.gov/media/145022/downloadDate: January 2021Date accessed: April 16, 2021Google already introduced other sectors. For example, Tesla continuously updates cars' autopilots basis feedback aggregated fleet approximately 500 000 vehicles.5Towards scienceTesla's deep at scale: using billions miles train networks.https://towardsdatascience.com/teslas-deep-learning-at-scale-7eed85b235d3Date: May 7, 2019Date This example shows create more advanced ML-based devices allow performance improvement. These improvements could include personalisation removal errors, would lead accurate outcomes. In systems, algorithms are trained specific dataset. Such often perform well similar but poorly scenarios rare training process.6Oren O Gersh BJ Bhatt DL imaging: switching radiographic pathological clinically meaningful endpoints.Lancet e486-e488Summary (46) collecting after authorisation, potentially support improved health However, poses risks need addressed. First, introduce errors. New subject reporting errors (eg, when wrong age diagnosis entered electronic records), inaccurate Second, might deteriorate if newly integrated biased.7Kaushal A Altman Langlotz Geographic distribution cohorts used algorithms.JAMA. 324: 1212-1213Crossref this type error, referred domain shift, an was developed both white black patients, were collected disproportionately eventually decrease accuracy outcome patients. Third, there risk information interfere what model (also catastrophic forgetting).8Kirkpatrick J Pascanu Rabinowitz N et al.Overcoming forgetting networks.Proc Natl Acad Sci USA. 2017; 114: 3521-3526Crossref (1403) Catastrophic might, worst case, overwrite model's previous knowledge deterioration performance.1Parisi 2019, published discussion paper, highlighting one benefits resides capability performance.9FDAProposed regulatory framework modifications (SaMD).https://www.fda.gov/files/medical%20devices/published/US-FDA-Artificial-Intelligence-and-Machine-Learning-Discussion-Paper.pdfDate 2021, issued plan facilitate innovation software.4FDAArtificial One major step regulation so-called predetermined change control authorisation process devices. shall aspects manufacturer intends (prespecifications) associated methodology being implement changes controlled manner manages patients (algorithm protocol).4FDAArtificial The provides clear pathway authorise inherent benefits, mean it's important takes cautious approach regulating At stage, most pressing issue determine prerequisites regarding how device. specify draft guidance.4FDAArtificial crucial principle ensure introduction reduction performance. To reach goal, should determine, during review process, device's further identifying addressing updating system, prespecifications approval relevant selected. unbiased. Otherwise, mistakes system. Flexibility appropriate some cases. radiology, images useful even correctness diagnoses ensured where missing altogether. image pretraining, before finetuned correct result better overall performance.10Esteva Robicquet Ramsundar B al.A guide healthcare.Nat Med. 25: 24-29Crossref (951) Post monitoring necessary, should, particular, degrade demographics practice changes. protocol determines applied achieve prespecifications. maintains improves unwanted algorithm, overwriting existing inference introducing standardised testing routines recommended.2Lee Testing current level maintained any successful. Additionally, check that, given probabilistic nature contemporary subgroups remains robust. Finally, mitigate hacking attacks updates, it connection WiFi Ethernet) encrypted, verified, secure. FDA's intended benefit addressed plan. All authors had final responsibility submit publication. KNV reports grants Swiss National Science Foundatioin (SNSF) writing submitted work Foundation Cancer Research outside work. SF ASK Arnold Ventures funding sources role content Comment publish it.